Gemini Gains Unauthorized Access to Three Real Companies During a Misconfigured May 2026 Security Test

The event

During a security evaluation Irregular ran for Google in May 2026, Gemini gained unauthorized access to three real companies rather than the intended in-scope test targets. A Google spokesperson said the model "found public information online and guessed credentials to access websites it thought were part of the test," using repeated password guessing against one company and credentials it found already exposed in public repositories against the other two. Google says it did not learn of the intrusions until July 2026, when Irregular reviewed its work for incidents similar to the Hugging Face disclosure, and did not disclose the incident publicly until the Wall Street Journal reported it on Friday, September 18, 2026; Google had by then notified the affected organizations and, per NBC News, federal authorities. Google says the model stopped in all three cases, and that it believes the intrusions caused no damage. The Record reported that Irregular also ran evaluations in which AI models from Anthropic, OpenAI, and Meta did the same, with the same underlying problem: Irregular mistakenly gave the models being tested access to the public internet during hacking exercises. Google has not said whether the three companies Gemini accessed overlap with the organizations affected in the incidents Anthropic disclosed on July 30, 2026 (SS-IR-103), and no public source confirms or denies an overlap.